Reflection-Driven Control Overview

• Reflection-Driven Control is an approach that elevates intermediate reflective processes to control mechanisms across various domains.
• It integrates self-reflection in language models, symmetry-based regularization in reinforcement learning, and direct physical manipulation in engineering systems.
• This paradigm enhances performance by allowing systems to inspect and adjust internal states, optimize trajectories, and engineer wave or molecular responses.

Reflection-driven control is a domain-dependent term for control schemes in which “reflection” is elevated from a passive consequence to an explicit control variable. In recent work, the term denotes at least three distinct families of mechanisms: internal self-reflection that revises reasoning in LLMs and agents, geometric or symmetry-based reflection used as a control prior in reinforcement learning and stochastic control, and direct manipulation of physical reflection channels in molecular, electromagnetic, optical, and acoustic systems. What unifies these usages is not a single formalism but a common design move: an intermediate reflective process is inserted between input and output so that behavior can be redirected by inspecting, aligning, or engineering reflected structure rather than merely pushing a feed-forward map harder (Zhu et al., 13 Jun 2025, Zhen et al., 22 May 2026, Christensen et al., 2023, Meyer et al., 2 Apr 2025, Asadchy et al., 2016).

1. Conceptual scope and recurring meanings

In the LLM and agent literature, reflection-driven control refers to explicit manipulation of self-reflective reasoning. Reflection is operationalized as revisiting, evaluating, critiquing, or revising intermediate reasoning, often signaled by tokens such as “wait,” and then either amplified or suppressed at inference time. In this setting, the control target is not the final token directly but a latent reflective state that changes the downstream trajectory of reasoning, action selection, or repair (Zhu et al., 13 Jun 2025, Yan et al., 16 Dec 2025, Chang et al., 23 Aug 2025, Wang et al., 22 Dec 2025).

In reinforcement learning and stochastic control, the phrase refers to more literal meanings of reflection. One line exploits reflection symmetry in state-action space, so that mirrored states and actions can be regularized to share value and policy structure. Another line studies reflected diffusions, where the controller chooses a domain and the process is pushed back by normal reflection at the boundary through local time. Here “reflection” is not introspection but either a symmetry group action or a singular boundary control mechanism (Zhen et al., 22 May 2026, Christensen et al., 2023).

In the physical sciences, reflection-driven control retains its wave- or transport-theoretic meaning. The controlled object may be Andreev reflection across a single-molecule junction, anomalous electromagnetic reflection from metasurfaces or metagratings, acoustic reflection from a metascreen, or reflected power of partially coherent waves under unitary input transformations. A recurrent theme is that high reflectance is not controlled by a scalar transparency knob alone; instead, one engineers modal alignment, orbital resonance, phase, or diffraction channels (Meyer et al., 2 Apr 2025, Li et al., 2024, Guo et al., 2024, Chen et al., 2024).

2. Representation-engineered self-reflection in LLMs and code agents

A central line of work treats self-reflection as a latent, steerable internal feature. “From Emergence to Control: Probing and Modulating Self-Reflection in LLMs” shows that self-reflection is not exclusive to RLVR-tuned models: in ordinary generation, pretrained Qwen2.5-1.5B reflects only about 0.6% of the time on MATH500, but reflection-inducing probing raises that rate to 18.6%. The same paper defines a self-reflection vector by a mean-difference statistic in activation space,

$$v^{(\ell)}=\mu^{(\ell)}_{\text{ref}}-\mu^{(\ell)}_{\text{nonref}},$$

and uses residual-stream steering to obtain bidirectional control. On DeepSeek-R1-1.5B, SR enhancement improves MATH-500 from 84.1 to 87.4 Pass@1, AIME 2024 from 29.2 to 33.5, and GPQA Diamond from 14.0 to 18.9, while SR suppression reduces length on MATH-500 from 4755 to 3716 tokens with Pass@1 changing from 84.1 to 83.2. The same study reports a peak at $\alpha=0.03$ with a 12% improvement in Pass@1 and explicitly identifies an over-reflection regime in which more reflection degrades performance (Zhu et al., 13 Jun 2025).

“ReflCtrl: Controlling LLM Reflection via Representation Engineering” shifts the unit of intervention from the whole sequence to the reasoning step. It segments chain-of-thought by \n\n, labels reflection onsets using cues such as “Wait” or “Let me think,” extracts layerwise mean-difference directions separately in attention and MLP streams, and applies additive steering only at step boundaries:

$$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$

Its main empirical claim is efficiency-oriented: many reflections are redundant, especially in stronger models. The paper reports up to 33.6% reasoning-token savings while preserving accuracy, and for QwQ-32B at $\lambda=-0.96$ it reduces GSM8k reasoning tokens from 1488.6 to 1006.7 with accuracy changing from 96.50% to 96.36%. It also links reflection directions to an internal uncertainty signal by showing that projection features on these directions predict answer correctness better than final-layer embeddings on several models (Yan et al., 16 Dec 2025).

“Unveiling the Latent Directions of Reflection in LLMs” studies a narrower situational-reflection setup with three prompt-conditioned regimes: no reflection, intrinsic reflection, and triggered reflection. It defines steering directions between these regimes at the appended instruction token and shows both instruction discovery and causal intervention. On GSM8k-adv, triggered reflection averages .397 on Qwen2.5-3B and .586 on Gemma3-4B, versus .051 and .147 for no reflection. The paper further reports that suppressing reflection is considerably easier than stimulating it: for Qwen2.5-3B, inhibition can drive Wait-prompt accuracy down near no-reflection levels. This weakens the common assumption that reflective behavior, once present, is robust under perturbation (Chang et al., 23 Aug 2025).

In secure code generation, “Reflection-Driven Control for Trustworthy Code Agents” turns reflection into a standardized control module. A lightweight self-checker first emits SAFE or UNSAFE; unsafe cases trigger retrieval from a dynamic memory $M_D$ of verified repair cases and a static memory $M_S$ of secure-coding knowledge, using a retrieval rule that falls back to $M_S$ when dynamic retrieval is too sparse or too dissimilar. Across eight CWE classes, Base+Reflex improves security rate for all four tested models: for example, gpt-4o rises from 85.7 to 95.0 and qwen3-coder-plus from 83.7 to 94.9, while token and runtime overhead remain modest. The reflective memory also evolves: retrieval success rises from 85% to 100% across five runs, and fallback usage drops from 15% to 0% (Wang et al., 22 Dec 2025).

3. Reflection as closed-loop deliberation in agents, robotics, driving, and visual generation

In embodied and tool-using agents, reflection-driven control is typically a closed-loop correction mechanism. “ReflAct: World-Grounded Decision Making in LLM Agents via Goal-State Reflection” argues that ReAct’s main weakness is a mis-specified reasoning objective: thoughts focus on the next move instead of the current agent state relative to the task goal. In a POMDP

$$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$

ReflAct replaces next-action-oriented thought with state-goal-oriented reflection. The reported effect is substantial: on 134 ALFWorld tasks with GPT-4o, ReflAct reaches a 93.3% success rate and surpasses ReAct by an average of 27.7%; the paper also reports that ReAct reduces average action entropy from 1.23 for NoThinking to 0.30, which helps when the thought is grounded and harms when it is not (Kim et al., 21 May 2025).

For real-world manipulation, “PhysReflect-VLA” wraps a pretrained VLA policy in a feasibility-and-reflection loop. A forward Feasibility Operator $\mathcal{F}_\theta$ predicts candidate state transitions, an inverse Action Explanation Operator $\mathcal{G}_\psi$ reconstructs the action that would explain the transition, and a consistency energy

$\alpha=0.03$0

ranks candidates. After execution, reflection is triggered when

$\alpha=0.03$1

and the resulting guidance token is appended to the instruction. On OpenVLA fine-tuning, the full system improves average success from 74.2% to 79.6%, a gain of 5.4 points; removing feasibility or reflection lowers performance, and removing cycle-consistency during training drops success to 70.2% (Yang et al., 25 Jun 2026).

Autonomous driving papers instantiate the same pattern with different control substrates. “ReflectDrive” discretizes future trajectories into token sequences and adds a safety-aware reflective loop: generate diverse goal-conditioned trajectories, identify the earliest unsafe waypoint, perform local discrete search around that token, fix the replacement as a safe anchor, and inpaint the surrounding segment. On NAVSIM, the reflective inference stage raises PDMS from 84.8 to 91.1, with especially large gains in drivable-area compliance and ego progress. “IRR-Drive” inserts an explicit multimodal reflection stage between an initial textual intention and the final trajectory by predicting future semantic BEV under the preliminary intent, then refining the intent before action output. Its adaptive reflection mechanism yields PDMS 91.3 on NAVSIM v1 and EPDMS 89.0 on NAVSIM v2, while runtime remains 1.70 s/sample compared with 1.46 s/sample for always non-reflective inference and 3.03 s/sample for always reflective inference (Li et al., 24 Sep 2025, Chen et al., 22 Jun 2026).

A related generative-control example appears in “VisionCreator-R1.” There, reflection is inserted into a UTPCR trajectory—Understanding, Thinking, Planning, Creation, Reflection—and used to inspect intermediate images before issuing corrective editing calls. The paper’s key claim is an optimization asymmetry: planning is reliably optimized by plan rewards, while reflection is harder because its effect is downstream and noisy. On multi-image tasks, the final VisionCreator-R1 reaches 0.700 versus Gemini2.5Pro at 0.649, supporting the view that reflection can function as a control primitive when grounded in intermediate outcomes rather than treated as free-form critique (Lai et al., 9 Mar 2026).

4. Symmetry reflection and reflected-domain control in reinforcement learning and stochastic control

In continuous-control RL, reflection-driven control can mean exploiting reflection symmetry of the underlying MDP rather than any form of self-evaluation. “Reflex: Reinforcement Learning with Reflection Symmetry Exploitation in State-Based Continuous Control” formalizes a $\alpha=0.03$2-invariant MDP,

$\alpha=0.03$3

with reflection group $\alpha=0.03$4 satisfying $\alpha=0.03$5 and invariance conditions

$\alpha=0.03$6

The paper distinguishes axial reflection, defined by $\alpha=0.03$7, from bilateral reflection, which combines axial transformation with left-right swapping. It proves $\alpha=0.03$8, $\alpha=0.03$9, and that an optimal policy can be chosen $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$0-equivariant. Reflex then implements these facts as symmetry regularizers in PPO and SAC. The reported gains are largest on bilateral locomotion tasks: PPO+Reflex improves Walker2d from $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$1 to $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$2, and SAC+Reflex improves Ant from $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$3 to $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$4 (Zhen et al., 22 May 2026).

A more literal reflected-control model appears in “Data-driven rules for multidimensional reflection problems.” The state follows a Langevin diffusion,

$$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$5

and the controller chooses a bounded domain $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$6; the reflected process is then

$$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$7

where $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$8 is the unit inward normal and $$z_{l,\text{intv}}^{\{\text{attn},\text{mlp}\}}=z_l^{\{\text{attn},\text{mlp}\}}+\lambda d_l^{\{\text{attn},\text{mlp}\}}.$$9 the boundary local time. The long-run control problem,

$\lambda=-0.96$0

is shown to equal the explicit domain functional

$\lambda=-0.96$1

This converts multidimensional singular control into shape optimization. The paper then proposes a gradient-descent method over strongly star-shaped polytope approximations and extends the setting to unknown dynamics, proving static regret at the minimax-optimal invariant-density estimation rate and sublinear online regret via an episodic exploration–exploitation algorithm (Christensen et al., 2023).

These two strands show that “reflection” in control theory is not merely metaphorical. In one case, reflection supplies a group action that constrains optimal value and policy structure; in the other, it is the control law itself, exerted by boundary local time. This suggests that reflection-driven control is broader than deliberative AI: it also names control strategies in which mirrored structure or reflected trajectories are the primary mathematical object.

5. Resonant control of Andreev reflection in single-molecule transport

In mesoscopic transport, “Control of Andreev Reflection via a Single-Molecule Orbital” uses the phrase in a strictly physical sense. The device is an $\lambda=-0.96$2-molecule-$\lambda=-0.96$3 junction in a scanning tunneling microscope, formed by an Au-coated W tip, a single phthalocyanine molecule, and a Pb(111) substrate. The central result is that subgap transport is not governed simply by barrier transparency, as in a bare BTK interface, but by steering a single molecular orbital through the Fermi level as the STM vacuum gap closes. The 2H-Pc molecule exhibits a movable near-$\lambda=-0.96$4 resonance, whereas Pc does not, and this difference controls whether resonant Andreev reflection appears (Meyer et al., 2 Apr 2025).

Experimentally, 2H-Pc shows a characteristic progression from an ordinary Pb BCS gap in the tunneling regime to a gap filled with spectral weight and then a pronounced zero-bias resonance at contact. The normal-state resonance has Frota line shape with full width at half maximum $\lambda=-0.96$5, yielding an estimated Kondo temperature of about $\lambda=-0.96$6, while the superconducting state requires a combined Frota + BTK description. The extracted effective BTK transmission $\lambda=-0.96$7 is nonmonotonic: it increases with tip approach, reaches about unity for $\lambda=-0.96$8, and then decreases again. By contrast, Pc remains largely BCS-like and its extracted transmission stays below about $\lambda=-0.96$9 even near $M_D$0. The paper therefore rejects the naive expectation that Andreev reflection must increase monotonically with conductance (Meyer et al., 2 Apr 2025).

The theoretical support is a one-dimensional $M_D$1-molecule-$M_D$2 Green-function model with a single adjustable molecular level $M_D$3, exchange $M_D$4, and couplings $M_D$5. The model omits Kondo correlations and is intended only to isolate orbital alignment. Its main result mirrors the experiment: as $M_D$6 approaches $M_D$7, the BCS gap fills in due to Andreev reflection; the signal is maximal when $M_D$8; once the level moves away, the Andreev feature weakens and the gap recovers. Here reflection-driven control means control of electron-to-hole conversion by resonance alignment of a single orbital, assisted by contact-induced hybridization broadening, rather than simple mechanical transparency (Meyer et al., 2 Apr 2025).

6. Electromagnetic, optical, acoustic, and coherence-based control of reflected waves

A second physical family concerns direct engineering of reflected wave channels. In metasurface theory, “Perfect control of reflection and refraction using spatially dispersive metasurfaces” establishes a foundational limitation: local passive lossless phase-gradient reflectors cannot realize same-polarization anomalous reflection into an arbitrary direction without parasitic scattering, because the incident and reflected waves interfere along the surface and create an oscillating normal Poynting flux. The paper shows that ideal same-polarization anomalous reflection requires strong spatial dispersion, interpreted as non-local lateral power channeling along the metasurface, whereas perfect orthogonal-polarization reflection can remain local and lossless (Asadchy et al., 2016).

Later work on metagratings shifts from continuous phase profiles to diffraction-channel engineering. “Functional control of anomalous reflection via engineered metagratings without polarization limitations” studies periodic air grooves on a PEC surface and chooses the grating period so that only the $M_D$9 and $M_S$0 reflected orders propagate. The design task is then to suppress the specular order $M_S$1 and route power to the retroreflected order $M_S$2. The paper derives explicit inverse formulas for TM and TE retroreflection and validates them by full-wave simulation at 2 GHz, showing cases of polarization-dependent retroreflection, dual retroreflection, and polarization-independent retroreflection controlled mainly by groove depth (Li et al., 2024).

Optical atomic-media papers implement reflection control through coherence and phase. “Phase control of transmission and reflection in a sample of duplicated two-level systems driven by a stationary control field” introduces a standing-wave control field and a weak orthogonally polarized probe, with an effective susceptibility

$M_S$3

so the reflected and transmitted probe factors become tunable by the control–probe relative phase $M_S$4 in the moderate-optical-depth regime. “Tunable superluminal reflection and transmission through a slab doped with Raman driven atoms” instead uses a four-level $M_S$5-type Raman medium, where varying the control-field strength $M_S$6 changes gain and dispersion enough to switch reflected and transmitted pulse peaks from subluminal to superluminal. In the reported examples, reflection changes from $M_S$7 at $M_S$8 to $M_S$9 at $M_S$0 for one slab-thickness family (Hashmi et al., 2024, Ali, 2024).

In acoustics, “Two-sided Acoustic Metascreen for Broadband and Individual Reflection and Transmission Control” makes reflection an independently programmable output channel. The two-sided acoustic metascreen uses three geometry parameters—$M_S$1, $M_S$2, and $M_S$3—to approximately decouple reflection phase, transmission phase, and reflection/transmission amplitude partition. The reported operating band is 4 kHz to 8 kHz, with near-$M_S$4 coverage for both reflection and transmission phase. The paper demonstrates reflection-side diffusion, $M_S$5 anomalous reflection, simultaneous transmission-side focusing, and two-sided 3D holography with a $M_S$6 panel (Chen et al., 2024).

A more abstract wave-theoretic treatment appears in “Unitary control of partially coherent waves. II. Transmission or reflection.” For an incident partially coherent state $M_S$7 and reflectance operator $M_S$8, the total reflected power attainable under arbitrary unitary input control is exactly

$M_S$9

The same paper gives exact criteria for partially coherent perfect reflection, $$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$0, and partially coherent zero reflection, $$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$1. It also proves that if $$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$2, then $$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$3: more coherent inputs admit wider reflection-control intervals (Guo et al., 2024).

7. Cross-cutting themes, misconceptions, and open problems

A recurring misconception is that reflection-driven control always means self-reflection in LLMs. The literature here shows a much broader landscape: internal self-reflection in reasoning systems, reflection symmetry in RL, singular boundary reflection in diffusions, and direct control of physical reflection channels in molecular transport and wave physics all use the term in technically different senses (Zhu et al., 13 Jun 2025, Zhen et al., 22 May 2026, Christensen et al., 2023, Meyer et al., 2 Apr 2025).

Another misconception is that “more reflection” is uniformly beneficial. Several papers explicitly reject this. Activation-steered LLMs exhibit over-reflection, where positive steering eventually degrades Pass@1; ReflCtrl finds that many reflections are redundant and can be suppressed with up to 33.6% token savings while preserving performance; VisionCreator-R1 distinguishes under-reflection, good reflection, and over-reflection; and in molecular transport, higher junction conductance does not imply stronger Andreev reflection once the resonant orbital moves away from $$\mathcal{M}=\langle \mathcal{U},\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{P},\mathcal{R}\rangle,$$4 (Zhu et al., 13 Jun 2025, Yan et al., 16 Dec 2025, Lai et al., 9 Mar 2026, Meyer et al., 2 Apr 2025).

The principal limitations are domain-specific but structurally similar. Representation-engineered LLM control requires internal activations and remains unavailable in closed APIs; reflection detectors are often keyword-based and model-specific; stronger positive steering can perturb useful computation (Yan et al., 16 Dec 2025, Chang et al., 23 Aug 2025). Embodied systems depend on the quality of their predictive world models or future BEV forecasts and currently incur nontrivial inference cost; IRR-Drive, for example, is still far from real-time deployment despite adaptive routing (Yang et al., 25 Jun 2026, Chen et al., 22 Jun 2026). Physical reflection-control systems often face narrowband, angle-specific, or non-locality constraints: ideal same-polarization anomalous reflection requires strong spatial dispersion, metagrating solutions are geometry- and angle-conditioned, and optical phase-control formulas are cleanest only in moderate-optical-depth regimes (Asadchy et al., 2016, Li et al., 2024, Hashmi et al., 2024).

Taken together, these works suggest a general pattern rather than a single recipe. Reflection-driven control inserts a structured reflective layer between proposal and outcome: a model can inspect its own latent state, an agent can compare state to goal, a robot can compare predicted and observed transitions, a controller can exploit mirrored dynamics, a diffusion can be confined by a chosen reflecting boundary, and a physical device can reallocate power among reflection channels. The common implication is that control often improves when reflection is made explicit, diagnosable, and steerable rather than treated as an incidental side effect (Kim et al., 21 May 2025, Li et al., 24 Sep 2025, Christensen et al., 2023, Guo et al., 2024).